Installation with anaconda cloud fails

[Hemantr05]

conda install pytorch3d -c pytorch3d

gives the below output:

Collecting package metadata (current_repodata.json): done
Solving environment: failed with initial frozen solve. Retrying with flexible solve.
Solving environment: failed with repodata from current_repodata.json, will retry with next repodata source.
Collecting package metadata (repodata.json): done
Solving environment: failed with initial frozen solve. Retrying with flexible solve.
Solving environment: |
Found conflicts! Looking for incompatible packages.
This can take several minutes. Press CTRL-C to abort.
failed

[bottler]

Which operating system are you on? Can you please paste the output of conda list? Thanks

[Hemantr05]

Which operating system are you on? Can you please paste the output of conda list? Thanks

Ubuntu 18.04 is the OS I'm on.

[bottler]

The conda environment is totally empty (as shown by conda list). I would expect, in that case, conda install -c pytorch pytorch torchvision cudatoolkit=10.0 to install python and other things, and that's what happens for me. Once this has failed, there is no chance of being able to install pytorch3d.

Does conda install python work? Do you have other conda environments where python is working? If not, then I think you have a problem with conda. Updating/reinstalling anaconda might help. For example, conda uses the pycosat library to work out what to install, which has some versions which were behaving strangely. It might be interesting to do conda activate (to go back to the base) then python -c 'import pycosat; pycosat.__version__'. If it is less than 0.6.3 then that might be it.

EDIT: I meant to say python -c 'import pycosat; print(pycosat.__version__)

[Hemantr05]

I have pytorch and torchvision installed in my environment. (If you see the first line of the screenshot)
And as you can see, cuda==9.1 was installed.

And yes, conda install python works

The issue might be due the cuda version.

[bottler]

I don't see pytorch and torchvision installed in the first line of the screenshot. I see a failed attempt to install them. Segmentation fault (core dumped) means there was a catastrophic error, and in this case I expect nothing has been done to the environment by this command. That's why I'm suspicious of your conda installation.

conda list is showing that there is nothing installed in the environment.

If you can install python, it might be worth trying to install the other packages one by one. I would start with conda install cudatoolkit=9.2.

It is true that cuda 9.1 which I think you have on your machine is too old for pytorch3d. I think this should not matter if you install a newer version of the cudatoolkit conda package (say 9.2 or 10.0), as long as your nvidia driver is new enough. I am less worried about the UnsatisfiableError than about the Segmentation Fault.

[Hemantr05]

The screenshots below show the output of conda list, and includes pytorch as well.

[bottler]

Thanks. That's (most of) the output of conda list in your base environment. It has a lot of packages, including pytorch 1.3.1 which is too old for pytorch3d. It makes sense that (like in your first message) it fails with Found conflicts! or similar.

In your second message I see part of an attempt to install in a different environment, called pytorch3d. What I said in my previous message will still apply to that environment (and the conda installation which governs it). conda list looks different there.

[Hemantr05]

Will check on it, and get back.

[nikhilaravi]

@Hemantr05 were you able to resolve this problem? If so, please comment with what you needed to do to resolve it so other people can also benefit! Then, please close this issue.

[raincrash]

Weirdly, I had the same issue as well on OSX 10.15.3. conda install on pytorch3d channel returns a PackagesNotFoundError. I'm running miniconda3 with conda version 4.8.2. Is there some dependecy of the build on cudatoolkit, because the machine I tried is CPU only. I could install it from source though.

[raincrash]

I am not entirely sure how conda packaging works, but do we need to have a CPU-only version here or set USE_CUDA flags while building?

[bottler]

@raincrash MacOS is a separate issue. There are no conda packages for pytorch3d on MacOS at the moment, all the conda packages are for CUDA use with linux.

It should be straightforward to install CPU-only from source on a mac with pip if you have pytorch, fvcore, and torchvision installed, and if you have some xcode developer tools, e.g. with MACOSX_DEPLOYMENT_TARGET=10.14 CC=clang CXX=clang++ pip install 'git+https://github.com/facebookresearch/pytorch3d.git'.

We are planning to release pip binary packages for mac soon, which will make this easier.

If you have any more questions about mac, please open a separate issue.

[Hemantr05]

@Hemantr05 were you able to resolve this problem? If so, please comment with what you needed to do to resolve it so other people can also benefit! Then, please close this issue.

No, @nikhilaravi , the installation with conda failed. I guess it's because of the cuda version (requirements specify the cuda version to be 9.2 and above)